position-tester results

	A	B	C	D	E	F	G	H	I	J	K	L	M	N	O	P	Q	R	S	T	U	Z	AA	AB	AC	AD	AE	AF	AG	AH
1	Network	10	64	128	256	512	1000	Top 2	Top 3	Policy Elo	Value Elo	Ordo Elo	dev-log	Position Set	Nodes	MinibatchSize	SmartPruningFactor	Threads	BackEnd	GPU
2	60001		7.11	7.46	8.19	10.61	11.94	19.72	25.59	1471		-1200	Starting with kld threshold of 12micronats. Continuing with absolute fpu rather than the relative used recently on T51+ - although using absolute 1 at root instead of -1.	Medium10k	1000	8	1	1	cudnn-fp16	RTX 2080
3	60002									1492		-982	Starting with 16k games per net rather than the 32k that T40 was using.	Easy20k	10	1	1	1	cudnn	GTX 1060
4	60003									1516		-852
5	60004									1521		-745
6	60005		10.10	10.18	11.03	12.75	14.11	21.88	27.94	1516		-703
7	60006									1502		-648
8	60007									1483		-592
9	60008									1493		-479
10	60009									1494		-407
11	60010	23.86	14.29	14.31	14.53	14.96	16.47	26.84	34.35	1513		-227
12	60011									1533		-174
13	60012									1532		-115
14	60013									1539		-108	Increasing kld threshold to reduce the nodes per move average down from its current ~1900 level (22 micronats). Will increase games per net to 16k so net frequency shouldn't increase. (60013)
15	60014									1549		5
16	60015		16.51	16.73	17.49	18.23	20.20	33.11	41.58	1545		72
17	60016									1570		75
18	60017									1581		118
19	60018									1613		260
20	60019									1652		255
21	60020	29.51	19.18	19.79	20.80	21.78	23.41	38.63	48.77	1645		374
22	60021									1676		446
23	60022		15.76	16.19	17.16	18.14	20.38	33.62	42.46	1556		24	An issue was found with fp16 backends in combination with networks trained using mask_legal_moves. Because the network has no incentive to give low values for illegal moves it was possible for it to output values large enough that softmax would normalize legal moves to actual 0. This showed up most obviously on fp16 due to the limited range of exponents, but could theoretically affect all cases. Therefore mask_legal_moves has been disabled for now. Before it is turned back on we will need to add support to the engine backends to mask legal moves in their own softmax calculations. (60022 first net with mask_legal_moves disabled.)
24	60023		16.97	17.08	18.16	18.93	20.76	35.27	44.72	1604		213
25	60024									1588		401
26	60025		18.40	19.10	20.07	21.38	23.23	39.37	49.91	1634		563
27	60026									1665		641
28	60027		20.30	20.99	22.08	24.35	26.30	43.17	53.37	1680		741	mask_legal_moves fix seems to have also reduced the average nodes per move (which makes sense) so dropping kld threshold to 18 micronats. (Partial effect 60027.)
29	60028		20.02	20.74	22.08	22.94	24.23	40.56	51.51	1695		731
30	60029									1704		839
31	60030	34.62	22.75	24.04	25.03	27.59	30.11	49.21	60.32	1720		1048
32	60031									1728		1030
33	60032									1743		1119	Dropping kld threshold to 16 (Partial effect 60032.)
34	60033									1753		1221
35	60034									1751		1217
36	60035		23.40	24.71	27.16	29.10	32.28	52.49	63.34	1766		1245
37	60036									1774		1287
38	60037									1780		1312
39	60038									1798		1368
40	60039									1809		1402
41	60040	36.29	24.37	26.21	28.35	31.44	35.06	55.02	66.42	1810		1394	kld threshold to 14 as the nodes per move continues to fall. (Partial effect: 60041.)
42	60041									1815		1432
43	60042									1817		1421
44	60043									1829		1478											3.76	5.00	5.63	5.14	4.17	1.96	1.06	<-- StdDev
45	60044									1821		1488											1	2	3	3	2	2	1	<-- Weights
46	60045		24.92	26.96	29.72	32.74	36.49	56.72	68.48	1839		1494											0.27	0.40	0.53	0.58	0.48	1.02	0.94
47	60046									1832		1518								Top 2 (or 3) is the percentage where the ICCF move appears in the top two (or three) PVs at 1000 nodes, even if the first PV does not match.
48	60047									1841		1530	kld threshold to 12 (a bit early, but I expect it may pass below target while I'm busy). (Partial effect: 60048.)	Benchmarks	64	128	256	512	1000	Top 2	Top 3	net	64	128	256	512	1024	top2	top3	Score
49	60048									1856		1584		40475	29.70	36.56	44.82	55.67	66.94	83.91	90.88	badgyal-6	30.45	36.06	42.94	51.47	61.27	81.13	88.85	271.42
50	60049									1860		1588		22154	34.72	43.40	52.55	61.04	69.90	86.65	93.08	54255	29.00	34.47	42.49	52.87	64.18	82.68	90.06	275.10
51	60050	37.23	25.88	28.28	31.72	35.01	38.64	60.22	71.67	1862		1595		50778	29.24	36.49	45.56	57.71	70.73	87.10	92.89	goodgyal-6	32.15	38.71	46.12	55.29	64.13	83.94	91.28	283.39
52	60051									1879		1609		LD2	30.77	39.12	50.13	62.56	75.65	89.27	94.01	40475	29.70	36.56	44.82	55.67	66.94	83.91	90.88	282.35
53	60052									1856		1640		591219	35.71	45.97	58.18	71.24	83.40	93.04	96.27	53316	29.51	36.78	45.32	56.77	69.87	86.05	92.24	288.17
54	60053									1870		1642		11248	41.43	52.41	64.58	76.30	86.23	94.70	97.36	50778	29.24	36.49	45.56	57.71	70.73	87.10	92.89	290.75
55	60054									1885		1637		32930	39.31	51.64	64.30	76.21	87.41	94.78	96.97	52377	30.53	37.82	47.49	59.56	71.78	87.13	92.83	294.22
56	60055		25.58	28.20	30.85	34.42	39.25	60.93	72.16	1891		1662		J20-460	37.30	49.71	63.14	76.05	87.53	95.53	97.90	22154	34.72	43.40	52.55	61.04	69.90	86.65	93.08	299.96
57	60056									1920		1689		42810	41.53	55.03	69.55	81.74	92.65	97.40	98.58	51458	30.33	37.53	47.40	59.35	72.25	87.17	92.78	294.10
58	60057									1898		1704		J13B.2-136	45.47	60.06	73.67	84.92	94.00	98.21	99.05	57344	30.67	38.03	48.56	60.60	72.60	87.23	92.82	296.00
59	60058									1896		1709		LS 15	50.86	66.07	78.68	88.83	95.20	98.46	99.26	57382	30.77	37.61	48.87	61.09	72.44	87.21	93.00	296.38
60	60059									1920		1742		384x30-t60-4155	61.66	73.63	84.20	91.42	96.03	98.89	99.51	Dark Qeen 2.3	38.49	46.88	55.97	64.90	73.70	89.46	94.80	312.75
61	60060	40.67	25.51	27.99	31.14	35.13	40.08	61.42	72.83	1917		1721		66513	63.74	75.82	85.21	91.77	95.81	98.94	99.57	40500	31.54	40.37	51.35	63.27	75.43	88.55	93.32	303.39
62	60061									1925		1750		Best T60	64.26	76.16	85.44	91.95	95.81	99.03	99.59	LD2	30.77	39.12	50.13	62.56	75.65	89.27	94.01	303.11
63	60062									1922		1727		Corresp. Net	66683	66526	66516	66517	66513	66512	66522	128x10-t60-2-2010	36.13	46.76	58.31	70.41	81.99	92.18	96.15	324.57
64	60063									1931		1772										128x10-t60-2-2300	36.29	46.29	58.26	70.72	82.29	92.41	96.00	324.81
65	60064									1930		1756				Score (at 1kn)						128x10-t60-2-2990	36.34	46.76	58.95	71.34	82.14	92.72	96.31	326.28
66	60065		26.50	28.76	32.62	37.54	43.58	64.95	75.84	1951		1786		Best net tested yet:	66513	95.81						128x10-t60-2-5300	36.70	46.61	57.97	71.25	82.65	93.22	96.51	326.69
67	60066									1954		1834										591219	35.71	45.97	58.18	71.24	83.40	93.04	96.27	326.22
68	60067									1960		1858										41250	35.31	46.98	60.20	73.51	85.14	94.28	96.97	331.68
69	60068									1964		1840										11248	41.43	52.41	64.58	76.30	86.23	94.70	97.36	340.76
70	60069									1971		1876										32930	39.31	51.64	64.30	76.21	87.41	94.78	96.97	339.97
71	60070	44.52	27.15	29.67	34.41	38.77	45.55	66.43	77.82	1962		1851										J20-460	37.30	49.71	63.14	76.05	87.53	95.53	97.90	339.65
72	60071									1966		1892										J64-180	41.14	53.59	67.08	79.60	89.42	96.16	98.24	348.27
73	60072									1980		1892										J64-210	41.91	55.07	68.04	80.75	90.37	96.56	98.50	351.36
74	60073									1972		1883										SV-256x20-1810	46.62	58.33	70.64	81.30	89.97	96.67	98.59	355.63
75	60074									1987		1906										LS10.2	42.00	55.34	69.24	81.52	91.80	97.43	98.71	354.35
76	60075		27.33	31.15	35.38	41.09	46.76	67.70	77.51	1983		1927										42810	41.53	55.03	69.55	81.74	92.65	97.40	98.58	354.65
77	60076									1998		1935										T40B.4-160	41.63	55.57	70.05	82.33	92.34	97.64	98.71	355.72
78	60077									2002		1936										SV-20b-t40-1425	45.17	58.12	71.42	83.16	92.13	97.15	98.65	358.24
79	60078									2012		1958										SV-20b-t40-1541	45.56	58.62	71.98	83.57	92.22	97.27	98.81	359.40
80	60079									2001		1958	kld threshold to 10. (Partial effect: 60080)									jj-9-swa-220	43.11	57.18	71.20	83.47	92.99	97.53	98.83	358.35
81	60080	45.33	28.19	31.58	36.18	41.92	49.27	69.81	80.40	2020		1985	swa_max_n to 5. (60080)									LS 12.1	44.36	57.58	72.40	84.37	93.32	97.92	99.00	360.73
82	60081									2032		1970										J13B.4-90	45.28	59.68	73.24	84.36	93.39	97.99	99.05	362.40
83	60082									2020		1993										J13-410	44.67	58.96	72.87	84.42	93.51	98.10	99.02	361.93
84	60083									2031		1993										J13B.3-160	45.45	59.58	73.47	84.89	93.79	98.14	99.05	363.19
85	60084									2033		2009										J13B.2-136	45.47	60.06	73.67	84.92	94.00	98.21	99.05	363.68
86	60085		28.22	32.34	36.52	42.09	49.85	71.06	81.36	2039		2002										LS 13	45.70	59.74	73.67	85.48	93.87	98.14	99.14	363.89
87	60086									2040		2016										J13B.2-188	45.90	60.10	73.84	85.18	93.93	98.15	99.09	364.00
88	60087									2034		2035										SV-30b-t60-2-1504	52.13	64.34	76.24	85.74	92.77	97.81	99.11	368.07
89	60088									2043		2020										384x30-t40-1207	48.39	62.44	75.26	85.94	93.59	98.04	98.93	366.37
90	60089									2054		2037										384x30-t40-1705	52.07	65.49	77.88	87.44	93.85	98.12	99.13	371.23
91	60090	47.39	28.16	32.19	37.11	43.60	51.55	72.49	82.60	2050		2038										SV-30b-t40-1573	51.14	64.81	77.11	87.28	94.02	98.46	99.21	370.71
92	60091									2046		2039										384x30-t60-2481	54.58	67.51	78.46	87.33	93.47	98.12	99.38	373.01
93	60092									2050		2039										LS 13.2	47.47	62.92	76.19	87.59	95.12	98.38	99.22	369.13
94	60093									2053		2053										jj-9-563	48.74	64.05	77.05	87.75	95.15	98.21	99.13	370.23
95	60094									2071		2065										LS 14	48.65	63.56	76.98	88.04	95.42	98.40	99.15	370.48
96	60095		28.58	32.62	38.53	45.29	53.16	74.29	83.79	2065		2065										jj-9-5941	49.66	64.50	77.80	88.19	95.53	98.54	99.26	371.95
97	60096									2067		2069										LS 14.1	50.06	64.22	78.14	88.23	95.65	98.60	99.28	372.29
98	60097									2075		2072										jj-9-6213	50.11	64.98	78.44	88.44	95.44	98.43	99.23	372.56
99	60098									2072		2079										jj-9-6238	50.23	65.45	78.62	88.51	95.55	98.51	99.27	373.09
100	60099									2088		2073										jj-9-6242	50.00	65.13	78.92	88.48	95.50	98.52	99.22	372.99